Classifying multi-agent cooperative behavior is a fundamental problem in various scientific and engineering domains. In team sports, many cooperative plays can be manually labelled by experts. However, it requires high labour costs and a large amount of unlabelled data is not utilised. This paper examines semi-supervised learning methods for the classification of strategic cooperative plays (called screen plays) in basketball using a smaller labelled dataset and a larger unlabelled dataset. We compared the classification performance of two basic semi-supervised learning methods: self-training and label-propagation. Results show that the classification performance of the semi-supervised learning approaches improved upon the conventional supervised approach (SVM: support vector machine) for minor types of screen-plays (flare, pin, back, cross, and hand-off screen). For the feature importance, we found that self-training obtained similar or higher Sharpley values than SVM. Our approach has the potential to reduce manual labelling costs for detecting various cooperative behaviors.